Multi-factory scheduling presents notable challenges when road congestion exhibits temporal variability, as transport lead times deviate from predefined production plans and schedules deemed feasible at the planning stage may become unfeasible upon dispatch initiation. This study, therefore, addresses the problem through an integrated pipeline of traffic prediction, traffic parameterization, and joint scheduling, which links traffic prediction with scheduling within a unified decision framework rather than treating them as two separate tasks. Link-speed data from the Q-Traffic dataset are utilized to train a Light Gradient Boosting Machine (LightGBM) short-horizon predictor; the predicted speeds are converted into a bounded congestion factor, and a route-wise aggregation method is applied to derive the traffic correction coefficient at the factory-customer-route-time level. Traffic information is incorporated into the scheduler via transport-time adjustments and delivery-related penalty terms, thereby incorporating traffic information without substantially changing the original scheduling structure. At the decision layer, each order is treated as an independent agent, and each scheduling action is defined by the ternary set (factory, route, dispatch time window). The decision layer is trained under the framework of Centralized Training with Decentralized Execution (CTDE) by adopting Multi-Agent Proximal Policy Optimization (MAPPO), a method that mitigates the weak local optima commonly generated by stage-wise planning strategies. Comprehensive experiments, including main performance tests, ablation studies, and robustness analyses, are conducted on small, medium, and large problem instances as well as on low, medium, and high congestion groups. Across all experimental settings, the integration of parameterization and MAPPO joint decision-making achieves a favorable trade-off among total cost, on-time delivery rate (OTD), and average delivery delay, while maintaining stable performance amid traffic congestion fluctuations. Predictor accuracy, reproducible experimental configurations, and summarized results are also reported to support the replication and reuse of the proposed method in traffic-aware production-transport coordination scenarios. In comparison with stage-wise planning and heuristic Earliest Due Date (EDD)-based baselines, the augmented MAPPO policy yields lower overall costs and higher OTD values on large-scale instances. Ablation results further demonstrate that it modulates the learned preferences for routes and time windows in scheduling decisions. Decentralized inference of the proposed model remains in the millisecond range across all tested problem scales, which validates the practicality of the model for rolling-horizon decision support in real-world operations.